Signs of AI writing: 12 tells that hold up, and 5 that prove nothing
The clearest signs of AI writing are patterns, not single words. Look for inflated significance ("a pivotal moment"), stock vocabulary like "delve", lists of three, "it's not just X, it's Y" contrasts, vague "experts say" attributions, chatbot leftovers such as "I hope this helps", and stray Markdown. One sign proves nothing. Several together are a reason to look closer.
Everyone now has a theory about how to spot AI text. Some of those theories hold up and some just catch careful human writers. This guide sorts them: the signs that are actually useful, the ones that aren't, and what to do when you think you've found AI writing, whether it's in someone else's text or your own.
What are the most common signs of AI writing?
These are the tells that show up again and again in chatbot output. They're listed roughly from most to least telling.
| Sign | What it looks like | How telling |
|---|---|---|
| 1. Chatbot leftovers | "Certainly! Here's…", "I hope this helps", "As of my last update", [insert name here] | Very: a person pasted a reply without reading it |
| 2. Markdown in plain text | **asterisks** around words, ### headings, bullet lists pasted into an email or form | Strong where the format doesn't use Markdown |
| 3. Inflated significance | "a pivotal moment", "an enduring legacy", "underscores its importance" | Strong in a cluster |
| 4. Stock vocabulary | delve, tapestry, testament, intricate, showcase, multifaceted | Strong in a cluster, weak alone |
| 5. "Not just X, it's Y" | "It's not just a tool, it's a movement." | Medium to strong when repeated |
| 6. The rule of three | "fast, flexible and reliable", three examples every time | Medium |
| 7. Vague attributions | "Experts say…", "Many believe…", "Critics have noted…" with no names | Medium |
| 8. Superficial "-ing" analysis | "…highlighting its role in shaping the community" | Medium |
| 9. Even sentence rhythm | Paragraph after paragraph of 18–25 word sentences | Medium |
| 10. Hedges and balance | "While X has benefits, it's important to consider Y" | Medium to weak |
| 11. Tidy closing summary | "In conclusion…", "Overall, X remains…", challenges-and-future-prospects endings | Weak to medium |
| 12. Sudden style shift | One section reads nothing like the writer's other work | Strong, if you know the writer |
The vocabulary sign is the one most people know, and it's real: after ChatGPT launched, words like "delves" jumped in frequency across millions of scientific abstracts (Kobak et al., 2024). Our list of AI words to avoid covers those words, with a plain swap for each.
What does Wikipedia's "Signs of AI writing" guide say?
Wikipedia's volunteer editors maintain a long field guide, Wikipedia:Signs of AI writing, built from real examples they've found in articles, drafts and talk pages. It's one of the most useful public resources on the subject. A few points from it are worth knowing even if you never edit Wikipedia:
- It's descriptive, not a rulebook. The page says it lists observations, not rules, and that text with these signs isn't automatically AI-written, because chatbots learned from human writing in the first place.
- It groups signs by type: content (inflated significance, vague attributions, formulaic conclusions), language (AI vocabulary, the rule of three, negative parallelisms like "not just X, but Y"), communication meant for the chatbot user, formatting and markup, and citation problems such as broken links or invalid DOIs.
- It warns against relying on detectors or on your gut. It cautions against relying solely on detection tools, which have non-trivial error rates, and says people overestimate their own ability to spot AI text.
- It lists signs that don't work, which brings us to the next section.
The guide is written for Wikipedia's purposes, so some signs (edit summaries, wikitext quirks) only apply there. The content and language sections carry over to most writing.
Which "signs of AI" prove nothing?
Some popular tells catch careful human writers far more often than they catch chatbots. Wikipedia's guide lists several of these as ineffective, and they line up with what we see in practice:
- Perfect grammar. Lots of people write clean, correct prose, especially professional writers and anyone who proofreads.
- Formal or "fancy" wording. Academic and legal writing was formal long before chatbots.
- Transition words on their own. One "additionally" means nothing. It's the stacking, every paragraph, that's a pattern.
- Prose that feels bland. Plenty of human writing is dull. "Robotic" is a feeling, not evidence.
- The em dash. Chatbots overused it for a while, but writers have always used it. Wikipedia's guide now files em-dash overuse under historical indicators, noting that newer models often avoid the dash entirely. Our guide on whether the em dash is a sign of AI covers this in detail.
Non-native English writers get caught by these more than anyone. Their writing tends to use a narrower, more predictable vocabulary, and research has shown detectors flag it as AI far more often than native writing (Liang et al., 2023).
What does AI writing look like in practice?
Here's a short paragraph of the kind chatbots produce, with the signs marked:
The town's annual festival stands as a testament to its enduring legacy [3, 4], showcasing a vibrant tapestry of local culture [4, 8]. It's not just a celebration; it's a cornerstone of community identity [5]. Experts note that events like these foster connection, resilience and pride [6, 7].
And a version a person might write:
The festival started in 1974 as a fundraiser for the fire station. It's still mostly run by the station's volunteers, and most years about half the town shows up for the chili contest.
The second version (written for this example) isn't fancier. It's specific: a date, a who, a detail you could check. Specifics are the strongest sign of human writing there is, and they're the thing chatbots can't supply on their own.
What are the signs of AI writing in essays and student work?
In essays the same signs show up in a particular shape: a thesis that restates the prompt, three body paragraphs of near-identical length, balanced "on the other hand" hedging on every claim, quotations or sources that don't exist, and a conclusion that summarises without adding anything. Invented citations are the strongest of these, because they can be checked. The rest are also common in rushed human essays, which is why signs alone shouldn't decide anything. If you're a student worried that your own essay reads this way, see how to make your essay not sound like AI.
What should you do if you spot signs of AI writing?
- Count, don't react. One sign is noise. Look for several different kinds in the same text.
- Check the checkable. Do the sources exist? Do the quotes say what the text claims? Invented references are much stronger evidence than word choice.
- Ask about the work. A writer can usually explain why they made a choice, where a fact came from, or what they cut. Wikipedia's guide lists this ability to explain one's own editorial choices as a sign of human writing.
- Don't treat a detector score as a verdict. Use it to decide where to look, not what happened.
If you're on the other side of this, accused when you wrote the text yourself, our guide on how to prove you didn't use AI covers the evidence that actually counts. And if you used Grammarly and you're unsure where you stand, see does Grammarly count as AI.
How do you check your own writing for these signs?
Read it aloud and look for the table above: the stock words, the threes, the "not just X" lines, the tidy summary at the end. Then run it through a detector before and after you edit, and treat the score as a pointer to the flattest paragraphs. Our guide on whether your writing sounds like AI has a five-minute self-check, and the main guide on how to make AI not sound like AI covers the fixes. The aim isn't to hide anything. It's to make sure the text sounds like you and says what you mean.
Frequently asked questions
What are the most common signs of AI writing?
Inflated significance, stock vocabulary like "delve", the rule of three, "not just X, it's Y" contrasts, vague "experts say" attributions, even rhythm, chatbot leftovers and stray Markdown.
What does Wikipedia say about signs of AI writing?
Its editors' field guide lists real examples, says the list is descriptive rather than rules, and warns that no single sign, detector or gut feeling is proof.
Is an em dash a sign of AI writing?
Not on its own. Wikipedia's guide now treats em-dash overuse as a historical indicator, since newer models often avoid the dash.
Is perfect grammar a sign of AI?
No. It's listed as an ineffective indicator: plenty of people write clean, correct prose.
Can you prove someone used AI from the writing alone?
No. Signs are reasons to look closer. Check sources, look at drafts, and let the writer explain their choices.